SLAM in the Dynamic Context

of Robot Soccer Games

Stefan Tasse, Matthias Hofmann, and Oliver Urbann
Robotics Research Institute
Section Information Technology
TU Dortmund University
44221 Dortmund, Germany
Abstract. This paper evaluates the benets of modeling the dynamic
environment of robot soccer games as a SLAM problem. Moving objects
such as other robots and the ball are not only tracked individually, but
modeled in a full state and used for localization at the same time. This is
described as an implementation of an ecient system capable of running
in real time on limited platforms such as the humanoid robot Nao, and
the systems benet is evaluated using real world experiments.
1 Introduction
For an autonomous robot, knowing its location and the current state of the envi-
ronment is essential as it represents the basis for planning and reactive behavior.
Generally, this task may involve dierent aspects: Localization in known envi-
ronments, mapping previously unknown environments, and tracking dynamic
elements therein. In most common literature [1], those aspects are handled sep-
arately. Either a robot is only expected to localize in a known structured envi-
ronment, leaving a choice among many existing solutions which mostly dier in
their suitability for dierent kinds of uncertainties occurring during the robots
operation, or no prior knowledge exists at all as assumed in simultaneous lo-
calization and mapping (SLAM) scenarios, which represents the other extreme.
Dynamic elements are generally ignored in both cases, either handled as noise
or explicitly ltered out [2, 3].
Real world applications, however, always consist of a mixture of all those as-
pects. Many features in the robots area of operations will be known beforehand,
either from oor plans for indoor, or aerial photographs for outdoor scenarios,
or from previous mapping eorts or other given specications. On the other
hand, such a priori information rarely covers all features which are of interest
for the localization task. Some might even be occluded due to recent changes
in the environment. Incorporating new information into a robots internal map
can therefore often improve its ability to localize precisely. Similarly, it can be
shown that explicitly incorporating dynamic features into the system improves
the estimation quality, both for the tracking result and the localization [4].
In the course of this paper the modeling of the full state in dynamic situations
will be covered. Section 2 gives an analysis and a brief overview over related work.
Section 3 describes the actual implementation of such a system with the practical
approximations necessary to reduce the problems complexity to a point where
it is applicable in real-time on limited embedded platforms, specically on the
Nao V3.3 robot. An evaluation is given in section 4 in the robot soccer context
of the RoboCup Standard Platform League (SPL), and the paper is concluded
in Section 5.
2 Modeling
This section addresses the dierent aspects of the problem in the context of
a team of cooperating autonomous agents acting in highly dynamic competi-
tive environments, therefore settling certain design choices as the basis for the
implementation of the system described in section 3.
2.1 Localization and Robot-centric Tracking
The computational complexity of common ltering approaches naturally in-
creases with the states dimensionality. Separating the estimation of the robots
own position and that of the positions of other surrounding elements is therefore
motivated from a performance standpoint. Moreover, tracking dynamic objects
by stationary observers is a widely explored problem. Those are the main rea-
sons why the tracking problem for autonomous robots is commonly done in a
robot-centric local coordinate system.
The modeling of the localization and tracking aspects as a unied global esti-
mation problem, however, has some advantages compared to simple tracking in
robot-centric local coordinates. The latter necessitates the update of all tracked
objects with odometry data, which e.g. for humanoid robots can be extremely un-
reliable. Those propagate nearly uncorrected into infrequently observed targets,
leading to signicant drift. At the same time, the separate robot pose estimation
already corrected part of the odometrys errors and prevented the same drift in
its estimate. This motivates the advantage of modeling dynamic objects in global
reference systems. Additionally, information about tracked objects can be ben-
ecial for the localization when modeled in a unied state. Even if the objects
motion uncertainty prevents their use for accurate localization, shared informa-
tion about those objects among a team of agents might still resolve multi-modal
or symmetrically ambiguous localization states.
This leads to a heterogeneous system which has some resemblance to the
SLAM problem, since new features are mapped and used for localization at the
same time. In previous literature pure localization and the full SLAM problem
have been mostly separated. Only recent publications began exploring the in-
termediate between those extremes by incorporating a-priori information into
systems otherwise formulated as SLAM, e.g. in [5] where a SLAM approach is
augmented by a-priori information in form of aerial images. Moving objects are
just considered to be obstructive in normal SLAM algorithms and either handled
as noise [2] or tracked in separate model to be ltered out of the SLAM input,
therefore without any direct positive eect on the SLAM output [3].
An alternative system providing the same characteristics as the one proposed
here has already been published in [4]. It describes an adaptation of the Fast-
SLAM concept to include a-priori information as well as newly mapped static
and dynamic features with dierent degrees of uncertainty in their recognition
processes and motion models. The approach presented here diers from the one
in [4] in its vast use of Kalman lters in all dierent stages of the system, whereas
the latter integrates the SLAM aspects by use of a Rao-Blackwellized particle
lter. Notably, the system presented here can run in real-time on a Nao robot in
parallel to all other modules necessary to participate in RoboCup SPL games.
2.2 Heterogeneous Information Sources
As stated so far, the proposed system should use information about previously
known and previously unknown, static and dynamic features, and incorporate
all those into a coherent estimate of both the robots own positions as well as the
potentially dynamic states of the other objects. This obviously implies various
dierent, and more importantly, heterogeneous information sources.
Distinctions can be made according to the characteristics of each feature,
whether it can be used for localization directly or needs to be mapped, too, either
as a static but previously unknown feature or a dynamic one including motion
updates. A features associated uncertainty can vary both with respect to the
reliability and precision of its observation and the inherent predictability of its
motion model. Simple in-animated objects for example may just follow physical
equations of motion. Other autonomous agents on the other hand may change
their intention and action unpredictably, while being harder to measure reliably
due to their more complex shape, varying silhouette and changing backgrounds.
Each such distinction oers the possibility to apply approximations without
losing too much precision in the estimation result. A more thorough analysis
on the implication for those heterogeneous information sources and possible ap-
proximations can be found in [4]. In the implementation described in this paper,
only a subset of those is employed.
The most relevant approximation in the context of this paper is the ag-
gregation of measurements to build local short-term models of each observation
type, thereby decreasing the uncertainty associated with the observed target and
allowing to lter false positives. Once suciently recognized such a short-term
model can be forwarded to the central estimation system as a meta-measurement
and deleted from the temporary local model. The deletion of such models is im-
portant to preserve the independence assumption between consecutive measure-
ments which is important for the Bayes lter concept. Insuciently validated
local hypotheses on the other hand can be pruned away without eecting a
negative inuence on the systems estimate. The short life span of those local
models, e.g. below one second, prevents odometry errors to accumulate, but of-
ten allows the integration for example of a series of image processing results to
obtain superior measurement quality.
2.3 Distributed Modeling
The sharing of information among a team of autonomous agents is especially
desirable in cases where single robots have a very limited eld of view and when
occlusion frequently occurs. The distribution of information can be done with two
conceptually dierent approaches. One approach can be classied as bottom-up
and distributes measurements between robots, which are subsequently handled
by common sensor fusion techniques. This is for example done in [6] and [7]. The
top-down approach as applied in [8] and [9] consists of merging the individual
robots world maps. The prerequisite for such map merging is that all poses
of participating robots need to be known, either in a consistent global coordi-
nate frame or relative to each other. A common implementation in exploration
scenarios is with uniquely identiable robots which initiate map merging when
observing each other, or when all robots are condently localized in the global
reference frame.
This latter approach would exclude poorly localized robots from map merg-
ing, however, those might also prot from the shared information, even speci-
cally to resolve their poor localization in case of symmetries. If the measurements
are distributed among robots, basically only each sending robot needs to be lo-
calized successfully in a global reference frame. An additional advantage is the
computational and architectural simplicity of observation distribution compared
to map merging, especially if observations are already aggregated in temporary
local models as described in section 2.2, which in case of reliable localization
can be distributed at the same time when used for the local integration into
the global world model. Note that this approach does not guarantee a globally
consistent model among all robots, since insuciently localized robots do not
send out information and therefore integrate more (but potentially also more un-
reliable) knowledge. The dierence between the models of well localized robots
however can be minimized by globally scheduling the exchange and integration
phases of the observation distribution [4].
3 Implementation
The objective now is the realization of the demands specied in section 2, namely
to model the robots surrounding environment in one unied model using the
information of a whole team of robots as input. This is hardly possible to imple-
ment as a real-time system on an embedded platform without applying measures
to decrease the computational complexity. The presented approach consists of
three stages, which will be covered in the following sections. The rst stage han-
dles the static map information to realize most (but not all) of the localization
problem, and is based on an algorithm which can perform as a very ecient
stand-alone localization [10]. Parallel to this runs a stage performing local per-
cept aggregation according to the temporary short-term models described in
section 2.2. Finally section 3.3 presents the integration of local and distributed
perceptions into a consistent global world model.
In the following, each observation of a feature is represented by the two angles
z = (
1
,
2
)
T
describing the direction in which the feature has been detected,
and a third angle
3
in case the feature has an identiable orientation relative
to the robot coordinate system. This is visualized in gure 1.
Fig. 1. Measurement of a feature expressed in horizon aligned observation angles.
3.1 Multi-model Kalman Localization
In contrary to the system described in [4], which is based on a particle lter
localization, the approach presented here bases on a multi-hypothesis Unscented
Kalman Filter (UKF) localization which has been presented in [10]. This utilizes
an approach to Gaussian mixture ltering which combines the accuracy of the
Kalman lter and the robustness of particle lters without sacricing computa-
tional eciency. This is done by pointing out similarities to particle ltering with
an extremely low number of particles, and bypassing critical approximations in
common Gaussian mixture algorithms.
Applying known techniques from both elds in a new combination results in a
multi-hypotheses Kalman lter which is superior to common Kalman lters in its
ability of fast re-localization in kidnapped robot scenarios and its representation
of multi-modal belief distributions, and which outperforms particle lters in
localization accuracy and computational eciency. The output of this system
is a set of robot pose hypotheses with corresponding covariances and likelihood
estimations. If this is used as a separate localization module, the most likely
hypothesis can be considered as the localizations result, and used as an input
for behavior decisions or further planning.
To use this in the context of a unied world model it is necessary to keep
track of the history of each hypothesis origin for fusion and spawning of new
hypotheses, and the change of the likelihoods among the set of hypotheses, which
corresponds to a re-localization event for example with a kidnapped robot or
after temporary localization loss caused by extreme odometry errors or collisions.
Otherwise each estimates change can be considered as a pre-ltered input for
the global estimation system. This input bears the characteristics, on the one
hand, of partially corrected odometry data, and on the other hand that one of
buered and pre-processed sensor data. In addition to this, integration of further
information, including the communicated observations of other robots, can aect
the pose estimates, so those changes need to be fed back into the localization
module. The following sections will address the integration into the global model
and the stochastic soundness of this.
3.2 Local Percept Aggregation
When building upon the UKF localization described above, the full state can not
be factorized as in FastSLAM, but needs to be expressed as a joint probability
function, as in the EKF-SLAM solution. The increase of estimation complexity
by the high-dimensional state is countered by aggregation of some of the image
processing results into temporary percept-buers as motivated in section 2.2
with the aforementioned advantages.
This is applied to full extend to the dynamic features, i.e. the ball and other
robots in a robot soccer scenario. Measurements of robots or the ball consist
of two angles z = (
1
,
2
)
T
as described above. The estimated state consists
of a 2-dimensional spacial component and a corresponding velocity component:
= (x, y, v
x
, v
y
)
T
. For the spacial component of the state

= (x, y)
T
and the
height of the robots camera r, equations 1 and 2 can be used to calculate the
sensor model in form of the observation matrix H as in equation 3. Note that
the velocity is not observable by processing a single camera image.
h
1
() = atan2(r, |

|) (1)
h
2
() = atan2(y, x) (2)
H =
_

rx
|

|
3
+r
2
|

|

ry
|

|
3
+r
2
|

|
0 0

y
x
2
+y
2
x
x
2
+y
2
0 0
_
(3)
For objects which are simply governed by the physical laws of motion, in-
stead of being motorized or controlled, the motion model for the control update
consists of a continuous motion slowed down by a friction factor k =
F
friction
m
as
the force generated by the friction divided by the mass of the object. Since the
state is modeled in local coordinates, the robots own motion, given by the trans-
lational and rotational odometry (
x
,
y
,

), also transforms the local estimate.

This results in the following time update for the velocity vector:
v
t
=
_

_
=:V
..
_
1 +
kt
|v
t1
|
_
(

) v
t1
for |v
t1
| |kt|
_
0 0
0 0
_
v
t1
else .
(4)
where () =
_
cos() sin()
sin() cos()
_
is the rotation around . The full time update
therefore predicts the state
t1
according to equation 5.

t
= g(
t1
) =
_
(

) t(

)
0 V
_

t1

_
_
_
_

y
0
0
_
_
_
_
(5)
This results in the Jacobi matrix G for the process update as the partial
derivatives of x,y,v
x
and v
y
at (x
t1
, y
t1
, v
x,t1
, v
y,t1
):
G =
_
(

) t(

)
0 M
_
(6)
M =
_

_
_
gvx
vx
gvx
vy
gvy
vx
gvy
vy
_
if |v
t1
| |kt|
(

) else
(7)
with
g
vx
v
x
=
_
1 +
k t
|v|
_
cos(

)
k t v
x
(cos(

)v
x
sin(

)v
y
)
|v|
3
(8)
g
vx
v
y
=
_
1 +
k t
|v|
_
sin(

)
k t v
y
(cos(

)v
x
sin(

)v
y
)
|v|
3
(9)
g
vy
v
x
=
_
1 +
k t
|v|
_
sin(

)
k t v
x
(sin(

)v
x
+ cos(

)v
y
)
|v|
3
(10)
g
vy
v
y
=
_
1 +
k t
|v|
_
cos(

)
k t v
y
(sin(

)v
x
+ cos(

)v
y
)
|v|
3
. (11)
Thus local models of dynamic objects in the robots environment can be
modeled using separate Kalman lters. In case of the unpredictability of the
motion of autonomous robots it is possible to neglect the estimation of their
velocity and apply high process noise instead.
The separate localization module described in section 3.1, in itself also a
buer integrating information from static, known world features into a localiza-
tion belief model, is used analogically to those percept-buers, but the state is
not deleted periodically after forwarding the belief to the SLAM part of the algo-
rithm. This localization reects part of the SLAM state, and changes to this part
of the SLAM state are fed back into the localization modules state. Thus the
virtual localization measurements used to update the SLAM state are basically
the innovation introduced by new static feature observations. Therefore those
measurements are still conditionally independent from previous measurements
given the current belief state, so the Markov assumption is not violated.
3.3 Local and Distributed Knowledge Integration
The state of the full model of the robots environment consists of its own pose
p
0
= (p
0,x
, p
0,y
, p
0,
)
T
, the poses of all cooperating robots (p
i
= (p
i,x
, p
i,y
, p
i,
)
T
with i 1, ..., n), and the states of the dynamic objects. While only a small
subset of cooperating robots or other elements are observed at the same time
and modeled according to section 3.2 in each time interval, they remain part
of the full model also during time intervals where these are not observed. It
is possible to dynamically shrink or expand the state vector if new unknown
robots are observed. Alternatively a separate mechanism could keep track of
active and inactive slots in the state vector by using time-to-live counters. This
latter approach has been chosen here to prevent frequent rescaling of both the
state vector and its covariance matrix.
The integration of the locally accumulated and the distributed information
into the model will be done in the process and sensor update. The own pose and
those of cooperating robots can be updated with the pose changes propagated
from the individual localization modules relative to the pose used for the last
update. The ball is updated using a motion model similar to the one in equa-
tion 5, but without the odometry related rotations due to the local coordinate
system. Other autonomous agents can either be updated according to the latest
velocity estimations, or just using an identity and appropriately high process
noise following the reasoning proposed in section 2.2.
The sensor update consists of two dierent kinds of observations. If a robot,
either the local robot itself or any of the communicating robots in the team,
has made observations of static world elements which have been used to update
the separate localization estimate in the rst stage (cp. section 3.1), then this
absolute pose estimate is used as a direct measurement of the corresponding
pose in the state vector, i.e. the measurement Jacobian is an identity in the
corresponding submatrix.
The other case is the observation of a dynamic feature by one of the robots in
the team. If the observed dynamic feature is a robot (without further identied
characteristics such as team markers etc.), this dynamic object may either be
any of other robots in the team, or one of a number of non-cooperating other
robots in the environment. In this case, the maximum likelihood correspondence
will be chosen to be updated, or a new model will be inserted or activated if
the other choices are too unlikely. The corresponding expected observation is
in a robot-relative euclidean coordinate system, since this is the format of the
local models distributed as aggregated percepts. It is expressed as a function
of the observed objects model (m
x
, m
y
, m
vx
, m
vy
) and its observers pose p
i
,
with i = 0 for local observations and i 1, ..., n communicated ones, which are
otherwise not distinguished any further.
The observation model is given by equations 12 and 13
h
mx,my
(p
i
) = (p
i,
)
_
(m
x
, m
y
)
T
(p
i,x
, p
i,y
)
T
_
(12)
h
mvx
,mvy
(p
i
) = (p
i,
)
_
m
vx
, m
vy
_
T
(13)
from which the corresponding entries in the measurement Jacobian can be calcu-
lated as in equation 14, with c

and s

short for cos p

i,
and sin p
i,
, respectively.
_
_
_
_
_
_
_
hmx
mx
hmx
my
hmx
mvx
hmx
mvy
hmx
pi,x
hmx
pi,y
hmx
p
i,
hmy
mx
hmy
my
hmy
mvx
hmy
mvy
hmy
pi,x
hmy
pi,y
hmy
p
i,
hmvx
mx
hmvx
my
hmvx
mvx
hmvx
mvy
hmvx
pi,x
hmvx
pi,y
hmvx
p
i,
hmvy
mx
hmvy
my
hmvy
mvx
hmvy
mvy
hmvy
pi,x
hmvy
pi,y
hmvy
p
i,
_
_
_
_
_
_
_
=
_
_
_
_
c

0 0 c

(m
x
p
i,x
) s

+ (m
y
p
i,y
) c

0 0 s

(m
x
p
i,x
) c

(m
y
p
i,y
) s

0 0 c

0 0 m
vx
s

+ m
vy
c

0 0 s

0 0 m
vx
c

m
vy
s

_
_
_
_
(14)
Re-localization events can be handled by resetting the corresponding state
variables and removing the covariances, i.e. setting all entries in the covariance
matrix in the rows and columns to zero. If such a previous mis-localization by
a team member resulted in modeled false positives, those will stay as isolated
features in the state for some time and will be deleted or inactivated after a
certain time without observation. This serves as a self-repair routine to remove
clutter from the environmental model, and to prevent the growth of the state by
the accumulation of models of such elements. The same is done if two models of
unknown features are decided to correspond to the same origin after a series of
observations, so that the information needs to be fused into the rst model and
the seconds needs to be deactivated. Alternatively it would be possible to keep
multiple environment models for each localization hypothesis, as done in [4].
4 Evaluation
The modeling process is complex and incorporates a multitude of dierent infor-
mation, so that a step by step illustration of the working principle is not prac-
tical. To evaluate the presented approach, a simulated situation rst illustrates
the theoretical possibilities and the qualitative eect in section 4.1, followed by
a quantitative analysis in soccer games using experiments with real robots in
section 4.2. Both setups use an SPL scenario as specied by the 2011 rules.
4.1 Qualitative Demonstration
Figure 2 illustrates a simple scenario in a simulated environment. The robots in a
team share their information for distributed cooperative modeling. Figure 2(b)
shows the resulting model with 2D covariance ellipses extracted from the full
state. In the following, one robot looks down and does not see any static eld
features any more, and both he and the ball are teleported to another location
on the eld (see gure 3). The use of distributed percepts and the modeling of
the own pose together with the ones of other robots and the ball position and
velocity allows the robot to not only correct its position, but also its orientation.
(a) Setup of the robots on the eld. (b) World model generated from local
and distributed information.
Fig. 2. Scenario with a team of robots looking around and sharing perception infor-
mation to cooperatively model their environment.
(a) Scenario after teleportation of ball and
downwards-looking robot.
(b) World model generated from local
and distributed information.
Fig. 3. Following the situation in gure 2, one robot looks down and only sees the ball
but no landmarks, and he and the ball are teleported. The shared information however
still allows for a correction of both position and orientation of the robot.
This simple experiment shows the potential usefulness of such a combined
modeling of a robots dynamic environment and its pose in it. RoboCup SPL
games contain periods where robots are chasing the ball, approaching it for pre-
cise positioning to shoot at the goal, or even dribbling it. During those periods
odometry errors are integrated into the robots localization if not countered by
frequently looking up at static eld features to correct the robots pose estima-
tion. If looking at the ball also allows the correction of those odometry errors,
especially the orientation, this is expected to be a clear advantage.
4.2 Quantitative Performance Evaluation
The articial situation created in the previous section just serves as an example
of how localization benets may be gained. To allow a quantitative evaluation
of the approachs performance, the perceptions of a robot have been recorded
during normal game situations with real robots on a regular SPL eld. Those
perceptions include the proprioception, i.e. odometry, orientation and joint an-
gle information, exteroception, i.e. perceptions of objects by means of image
processing, as well as the distributed local models of other cooperating robots
running the same code, and ground truth information provided by a camera
system mounted above the eld.
This set of input information is then processed by two dierent module con-
gurations. One is the conguration described in section 3. The second uses
the same localization but a simpler module for cooperative tracking of dynamic
objects without any feedback into the localization, and has been used to win
the second place at RoboCup 2011. This experiment is not set up to show that
the localization works, since both solutions are based on the same competitive
solution for the localization problem with all features described in [10], but to
evaluate the additional benet gained by unied modeling of the full state.
A rst evaluation of several recorded situations did not show any conclu-
sive results, meaning the positive and negative eects of the full state modeling
equaled out most of the time, in a low percentage of cases the full system even
showed a slightly decreased localization quality. Closer evaluation showed that
the currently used visual robot recognition provides too much uncertainty or
even uncorrected systematical errors, such as in the distance estimation, to be
benecial for the localization.
A second conguration of the system, which ignores the robot perceptions for
the modeling of the robots own position, but still uses the much more precise ball
perceptions, showed the expected results. As can be seen in the representative
extract visualized in gure 4, the proposed system provides benecial information
for the robots own localization most of the time. A direct comparison of the
localization quality of both systems shows that the robot pose translation errors
for the full system model are below 25 cm in 83% of the time, and only 72% of the
time for the unassisted underlying localization module, and the average errors
over the whole experiment are 166 mm compared to 213 mm. However, note that
with this second conguration of the system, in the teleportation experiment the
robots orientation could not be recovered as easily as described in section 4.1.
Fig. 4. Dierence of translation errors of the two described systems. Negative values
mean larger errors of the unassisted localization compared to modeling the full state.
5 Conclusion
This paper presents the advantages of modeling the full environment state es-
timation as compared to only localizing in said environment. A competitive
stand-alone localization module is extended to perform as a full state model,
and the additional gain in localization performance is evaluated both in a simu-
lated situation as well as in several real world experiments with multiple robots
and ground truth provided by an external camera system. While the robot per-
ception in the current vision system is not good enough to benet from using
temporary opponent models as additional features for localization, usage of the
ball as a dynamic feature signicantly improves the localization quality.
An additional advantage of estimating the full state in a cooperative modeling
approach is the existence of a single model which contains all information in a
globally consistent way. This renders the switching between local tracking of the
ball and a global team ball model obsolete, for example, and therefore simplies
behavior specication.
References
1. Thrun, S., Burgard, W., Fox, D.: Probabilistic Robotics (Intelligent Robotics and
Autonomous Agents). The MIT Press (2005)
2. H ahnel, D., Schulz, D., Burgard, W.: Map building with mobile robots in populated
environments. In: Proc. of the IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS). (2002)
3. Schulz, D., Burgard, W., Fox, D., Cremers, A.B.: Tracking multiple moving objects
with a mobile robot. Computer Vision and Pattern Recognition, IEEE Computer
Society Conference on 1 (2001) 371
4. Czarnetzki, S., Rohde, C.: Handling heterogeneous information sources for multi-
robot sensor fusion. In: Proceedings of the 2010 IEEE International Conference on
Multisensor Fusion and Integration for Intelligent Systems (MFI 2010), Salt Lake
City, Utah (September 2010) 133 138
5. K ummerle, R., Steder, B., Dornhege, C., Kleiner, A., Grisetti, G., Burgard, W.:
Large scale graph-based SLAM using aerial images as prior information. In: Pro-
ceedings of Robotics: Science and Systems (RSS), Seattle, WA, USA (June 2009)
6. Stroupe, A., Matrin, M., Balch, T.: Distributed sensor fusion for object position
estimation by multi-robot systems. In: Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA), 2001. (2001)
7. Dietl, M., Gutmann, J.S., Nebel, B.: Cooperative Sensing in Dynamic Environ-
ments. In: Proceedings of the IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS). (2001) 17061713
8. Howard, A.: Multi-robot simultaneous localization and mapping using particle
lters. The International Journal of Robotics Research 25(12) (2006) 12431256
9. Zhou, X.S., Roumeliotis, S.I.: Multi-robot SLAM with unknown initial correspon-
dence: The robot rendezvous case. In: Proceedings of IEEE International Confer-
ence on Intelligent Robots and Systems (IROS), Beijing, China (2006) 17851792
10. Jochmann, G., Kerner, S., Tasse, S., Urbann, O.: Ecient multi-hypotheses un-
scented kalman ltering for robust localization. In R ofer, T., Mayer, N.M., Savage,
J., Saranli, U., eds.: RoboCup 2011: Robot Soccer World Cup XV. Lecture Notes
in Computer Science. Springer Berlin / Heidelberg (2012) to appear